Based On High-order Image Feature Descriptor Of The Image Retrieval Method

In recent years, with the rapid development of multimedia technology andnetwork technology, more and more digital images began flooding people’s lives.How to extract useful information from these vast amounts of image information isbecoming one of the most hot research subjects. In image retrieval developmentprocess, there has been many retrieval algorithms that are using the physically visualfeatures, such as color, texture, gray, shape and so on. These features are all low-levelvisual features of the physical image, although they are able to describe the basiccontent of the image, but when it comes to image deep semantic, they becomepowerless, so how to describe the image more properly becomes one of thedifficulties to be resolved for image retrieval.Bag of Words (BOW) model originated in from the document retrieval system.With its simple and effective advantages, it has been widely applied. Its core idea is touse a separate set of words to describe a document. Computer vision researchers try toapply the same thinking to the field of image processing and recognition, which madethe Bag of Visual Word (BOV) model. BOV model-based image retrieval technologyhas been widely used in image retrieval. On the basis of BOV model, we made aseries of studies, including several aspects as follows:(1) Since Hellinger kernel approach can get better results than the Euclideandistance, in this paper, instead of using Euclidean distance, we adoptive Hellingerkernel to measure the distance between SIFT descriptors. Through transform SIFTfeature descriptor into RootSift, the scalar product of RootSift is equivalent tocomputing the Hellinger kernel. This change is simple to implement in just a few linesof code, and it does not require any additional storage space.(2) The traditional weight calculation method is tf.idf, in order to get a betterperformance, we have introduced three other weight calculation methods, namely tf.χ2,tf.ig and tf.rf. Compared the effect of four weight calculation methods, weobtained an optimum weight calculation methods.(3) Construct higher-order image feature descriptors of an image. A n-orderimage feature descriptors is defined as n visual words in a certain spatial layout. Whenn is bigger than1, we call it high-order spatial features. During the calculation process,we use kernel function to reduce the calculation amounts.